Backporch
Judgement Lab · where the telling goes wrong

Judgement Lab

A reference tells you what a method is. This is the other half — instruments for the part a search box can't teach: where sound statistics gets told dishonestly, and where I build my own things to see. Move the sliders; the deception (and the topology) is live.

The persistent-homology explorer below is the same instrument derived from first principles in what is a shape made of? — points grow into cells, and the holes are counted as live Betti numbers.

Where the telling goes wrong

The math below is exactly right. Every lie lives in what gets shown and what gets left out — the gap between a sound result and an honest telling.

Torture the data — p-hacking

p = 0.0510
tests run 20“significant” by chance 1P(≥1 false positive) 0.642

Every dot is a test on pure noise — there is no real effect anywhere here. Under the null, p-values are uniform, so some land below 0.05 on their own. With 20 tries the chance of at least one “discovery” is 64%. The statistics is flawless. The lie is reporting the amber dot and never mentioning the 19 grey ones.

Same data, opposite story — Simpson's paradox

overall slope -0.666group A 0.774group B 0.552

Both groups trend upward (positive slopes), yet pull them apart and the overall line tips the other way. Same points, opposite stories. Nobody falsified anything — the lie is choosing whether to show the amber line or the two dashed ones.

Same statistics, different worlds — the Datasaurus

mean(x) 54.26mean(y) 47.83sd(x) 16.76sd(y) 26.93

Flip between the six. The scatter changes completely — yet the means and standard deviations don't move; they're identical to two decimals (≈ 54.26, 47.83, 16.76, 26.93) by construction. This is the Datasaurus (Alberto Cairo; Justin Matejka & George Fitzmaurice, 2017), the heir to Anscombe's quartet. The moral every analyst relearns: a summary can hide anything — including a dinosaur. A dashboard that reports only means is a place lies hide. Always plot the data.

Same GPA, different students — the transcript Datasaurus

cumulative GPA 2.804.00term →
cumulative GPA 2.80trend / term +0.28volatility (sd) 0.65

Started at 1.8, finished at 3.6 — someone who figured it out. The trend is the story, and probably the best bet of the five. A GPA cutoff would rank them even with the student who fell apart.

Five students, one GPA: the average is identical — 2.80 for every one — but the trend (teal) and the consistency are not, and those are the features a real decision needs. The mean is where you start asking, never where you stop. More in The average hides the student.

A 1-credit A is not a 4-credit A — credit weighting

Organic ChemistryA · 4.04 cr
Linear AlgebraA- · 3.74 cr
Intro SeminarC · 2.01 cr
Studio ElectiveB · 3.02 cr
Friday LabC- · 1.71 cr
arithmetic mean of grades 2.88 (ignores credits)GPA, credit-weighted 3.38

The registrar computes the weighted one — every grade counts in proportion to its credit hours. Drag the credits: the same five grades give a GPA anywhere from ~2.3 to ~3.5 depending only on where the A's sit. Put your A's in the four-credit courses and your low marks in the one-credit lab and you look like a different student — same letters, different weight. An unweighted average treats a Friday lab like a capstone.

How you'd actually decide — a holistic, weighted model

Climbed
88
The dip
67
Steady
64
Volatile
27
Collapsed
9

All five students have the same 2.80 GPA, so GPA itself can't rank them — it's constant. A holistic model scores what's left: the trend, the recent form, and the consistency. Move the weights and the order reshuffles — because there is no single right answer. That's the whole point: a real decision is an optimization over many weighted signals, tuned to what you actually care about — not a threshold on one mean. How do you decide? You decide what to weigh.

Something of my own — persistent homology

Not a borrowed demo. This is the topology I actually work in, made playable: a point cloud, a growing scale, and the holes that persist — computed live, Betti numbers and barcode and all.

1 · The space — your data, and the complex growing at scale ε
b₀ 1piecesb₁ 1holes (loops)
point · edge (two points within ε) · triangle (three within ε) · ◯ the ε-ball
2 · Persistence barcode — what survives as ε growsH₀ pieces H₁ holes
ε = 0.601.0782.155 ε →
…or as a persistence diagram — the farther above the line, the more real
born (ε) →dies
How to read it
Where this is actually used
Sensor networks
A persistent H₁ hole is a coverage gap no average would reveal.
Molecules & materials
Voids and channels — binding pockets, pores in an electrode — are H₁/H₂ features.
Time series
A loop in a delay embedding is a genuine cycle — periodicity a Fourier peak can miss.
Neuroscience & networks
Cycles in brain or contact graphs carry structure node-and-edge counts don't.
Images & shape
Counting holes and components classifies texture, porosity, and form.

Drag a point, click empty space to add one, double-click to delete — it all recomputes on your cloud. Real persistence, computed live: union-find for b₀, a GF(2) boundary-matrix rank for b₁.

Where this points: knots are made of cords with crossings, and the configuration of a tangle is exactly the kind of point cloud this instrument was built to read. The same diagram has a second, algebraic shadow — the Jones polynomial, computed live next door — so a knot can be measured two ways at once: by its topology here and its polynomial there. That pairing — persistence on knotted cord, against an invariant that can't be fooled by how it's drawn — is the thread I keep pulling.

See also — Contested claims: inference under weak signal and bias, worked through the parapsychology literature.